Asynchronous longitudinal designs have been widely used in studies on aging and Alzheimer's disease (AD). In these decades-long studies, as longitudinal variables were routinely collected from different platforms (e.g. food frequency questionnaires, lab results and clinical or psychological tests), the obtained data were not synchronized in time even for the same subject. As part of the motivating Normative Aging Study (NAS), it was of substantial interest to address whether dietary intake of niacin was associated with incident AD and cognitive decline. In the study, four cognitive tests were administered to participants in a span of 6 years after participation, and nutrient intake was determined by food frequency questionnaires distributed at the baseline and between adjacent cognitive tests. As a result, the exposure variables, including nutrient intake, and the outcomes of cognitive tests, were measured asynchronously. Naive analyses ignoring the mistiming issue will lead to biased results and wrong conclusion. In addition, the development of aging-associated diseases tends to differ in various (latent) subpopulations, where associations between disease progression and biomarkers and environmental exposures vary substantially across subpopulations. Subgroup analyses in aging populations are crucial for developing better strategies of preventing, treating and managing aging-associated diseases, such as AD. Accounting for the peculiar nature of the NAS data, the proposed methods will help address the key question on the association between dietary intake of niacin and incident AD in the NAS. We will also examine other biomarkers and environmental exposures associated with AD. The proposed subgroup analysis will also help identify latent subgroups within the aging population and discern subgroup-specific risk factors leading to cognitive impairment. Unlike the existing work relying on restrictive model assumptions, the proposed methods are fully nonparametric, data adaptive and can be combined with state-of-the-art model selection methods to choose relevant risk factors for AD from hundreds of variables that might be time varying and asynchronous with the response. Methods and theory developed in this project will be widely applicable to other longitudinal studies. We will also develop new computation tools and software packages that are scalable for large longitudinal data sets with long followups.